The discovery of druggable and structurally distinct allosteric sites across various protein classes has introduced new avenues for small molecules to modulate protein activity and, hence, cellular functions. Ligands that target allosteric sites may provide advantages like enhanced selectivity and often exhibit the possibility of targeting existing drug-resistant mutations. However, recent deep learning approaches show limited effectiveness in predicting allosteric sites, as demonstrated in the present study. We compare the performance of two deep learning methods, PUResNetV2.0 and VNEGNN, with Fpocket, a traditional geometry-based method and P2Rank, a geometry and machine learning ensemble approach.